[E-Daily Reporter Lee So-Hyun ] Naver (NAVER(035420)) is expanding its search service into an agent-based service that bridges the gap from information discovery to action, spearheaded by its next-generation “AI Tab.” Rather than competing with global Big Tech companies on the performance of general-purpose large language models (LLMs), Naver has focused on combining its own service assets—such as Search, Shopping, Places, and Reservations—with AI to help users complete the tasks they actually want to accomplish.
On the 2nd, Naver held a “Tech Deep Talk” session at Naver D2SF in Gangnam-gu, Seoul, under the theme “From Exploration to Execution: Naver AI Search Powered by Next-Generation AI Technology,” where it unveiled the core technologies behind the AI Tab. The three core technological pillars are: △ a “product-native LLM” optimized for the AI Tab; △ “harness engineering,” which ensures the AI performs services reliably; and △ “multimodal” technology based on Smart Lens.
AI Search Optimization… Twice the Speed and a 30 Percentage Point Reduction in Hallucination Compared to Previous Versions
The most notable distinguishing feature is that it is designed so the AI does not pretend to know what it does not know. Amid ongoing criticism of the “hallucination” problem—where generative AI plausibly provides answers that differ from reality—Naver emphasized that it has trained the AI tab not to fabricate arbitrary answers to ambiguous questions, but rather to ask for clarification.
Lee Ki-chang, Director of Hyperscale AI Models at Naver Cloud, explained, “While the existing HyperClova X (HCX) was a general-purpose LLM with broad knowledge and reasoning capabilities, the next-generation product-native LLM model focuses on selecting the right tools within the context of long conversations and completing the tasks users want from start to finish.” He added, “Our goal is not to rank first in every benchmark, but to create a model that performs best during real-world service moments when users search, purchase, and make reservations.”
Naver more than doubled the computing resources allocated to reinforcement learning compared to the existing HCX model. In particular, for questions it couldn’t answer, the company applied “Clarify RL,” a reinforcement learning method that provides rewards only when the model asks follow-up questions. Director Lee said, “The previous model tended to provide plausible answers even when the question was ambiguous,” adding, “We trained it to recognize that asking follow-up questions is beneficial when faced with ambiguous queries.” Naver reported that this approach reduced hallucination rates by up to 30 percentage points compared to the previous model.
Naver is also collaborating with global expert groups, such as the NVIDIA Nemotron Consortium, to further refine its models. The executive said, “We’re not just passively adopting external technologies; we’re jointly working on architecture design, training metrics, and diagnostics and analysis.” Performance is managed by dividing it into service, basic, and specialized capabilities. When the average of global models of the same class is set at 100, Naver’s service capability—which measures service quality—scored 108 points; basic capabilities, such as following instructions and invoking tools, scored 104 points; and specialized capabilities, such as solving doctoral-level scientific problems, scored 97 points. The executive explained, “We aim to develop a model that establishes a lead in service capability while narrowing the gap with the global best in specialized capabilities.”
Advanced through 27 Years of Search Infrastructure and Data/Service Assets
The AI Tab is not merely a chatbot-style search that provides answers based on what an LLM knows; rather, it is closer to an action-oriented search that enhances the reliability of answers by connecting to internal services—such as Naver Search, Blog, Cafe, Places, and Shopping—and the latest data. Han Seung-kyun, Leader of Naver’s AI Search Services, explained, “Having a single good LLM doesn’t guarantee a well-designed AI service.” He added, “In actual service deployment, we don’t rely solely on the model’s capabilities; instead, we utilize Naver Blog, Search, and various databases to provide the latest information in real time, enabling the system to generate more reliable answers based on that data.”
The core technology Han emphasized is “Harness Engineering,” which is the “know-how” that ensures the AI functions properly within the service. If a user asks, “I’m planning a company dinner in Jeongja-dong—find me a restaurant with good parking and available reservations,” the AI tab uses integrated search and Places search to find candidates, analyzes reviews to verify parking convenience, and then calls the reservation API to check availability for tonight. It then provides the answer along with actionable cards for reservations and directions.
To address cost and speed issues, Naver also implemented a division-of-labor structure that combines small language models (SLMs) tailored to specific roles. Instead of having a single large model handle everything, this approach uses specialized models for each stage—such as question classification, context summarization, tool invocation, and response generation. As a result, Naver reduced the operating costs of some AI Tab components by up to three times compared to the previous HCX-based structure and improved response speed by up to two times.
Visual Expansion of AI Search via Smart Lens, Prominently Placed in the Naver Search Bar
Multimodal technology represents the next direction for the AI Tab’s expansion. Yoon Sang-doo, Head of Naver’s Future AI Center, stated, “The visual search technology Naver has accumulated over nearly a decade serves as the visual foundation for AI agents to understand the world,” adding, “In the future, Naver’s AI agents will evolve to understand user intent not only through text but also through images, and connect that understanding to actual actions.” For example, if a user shows a photo or video and asks, “Book a place with this kind of atmosphere,” the AI will understand both the atmosphere of the space in the image and the text-based conditions, then connect the user to a places or booking service.
The AI Tab, which launched in beta for membership users last April, surpassed 3 million cumulative users in just one month. Since its official launch for all users at the end of June, the number of users has reportedly increased by 3 to 4 times compared to the beta period. The company has also outlined specific plans for future service expansion. In July—during the third quarter—Naver will create smoother integrations between the AI Tab and AI Briefing, as well as between the AI Tab and Smart Lens. In August, it will integrate real estate services into the AI Tab and introduce specialized agents in the Whale browser. Later this year, it plans to launch a health agent that provides health information based on verified sources such as academic papers and conferences.
The company took a cautious stance regarding the potential monetization of the AI Tab and usage limits. Naver stated that it currently has no plans to impose usage limits on the AI Tab. Regarding the introduction of ads in the AI Tab, the company said, “We currently believe that the reliability of answers and user trust are more important, so we have no plans to run ads.” However, it explained that the cost burden resulting from increased search volume and the service’s operational model are areas that will require future business decisions.
The competitive edge Naver highlights for the AI Tab is AI that has been validated through a search service used by the entire nation. The company explained that publicly launching the AI Tab on Naver—which receives tens of millions of visitors and processes hundreds of millions of searches daily—would be difficult without confidence in its quality, traffic, and cost-efficiency. Kim Sang-beom, Head of Naver’s Search Platform Division, emphasized, “This is virtually the first time a local company—rather than a global Big Tech firm—has provided a full-fledged conversational generative AI search service to the entire nation,” adding, “We will evolve Naver Search into an AI search experience that bridges the gap between exploration and action.”
